What is it about?
How to solve the data noise problem when using diffusion modeling for social information recommendation? The effect of some existing self-supervised learning methods is still limited. To address this, HKU's Data Intelligence Lab proposes a new project, RecDiff, a new diffusion model-based recommendation framework that better captures users' latent preferences and interests to generate personalized, tailored recommendations.
Featured Image
Photo by Alexander Shatov on Unsplash
Why is it important?
We propose a social diffusion model, called RecDiff, which combines the distinguishing capabilities of generative diffusion models with effective denoisingbased training objectives. Our RecDiff model excels in capturing data distributions and filtering noise in social graph structures, enabling accurate modeling of user similarities based on their preferences. Additionally, we employ a refined noise diffusion and removal process, allowing our social recommender to effectively handle various types of connection noise.
Perspectives
I hope that this article will contribute to the field of social recommendation.RecDiff is designed to be efficient and simple by employing a multi-step noise propagation and elimination training approach, running in hidden space and utilizing encoded user representations. By training the model in different diffusion steps, RecDiff demonstrates excellent noise handling capabilities and is able to effectively deal with the effects of various types of noise. We hope to communicate with more like-minded researchers!
Zongwei Li
University of Hong Kong
Read the Original
This page is a summary of: RecDiff: Diffusion Model for Social Recommendation, October 2024, ACM (Association for Computing Machinery),
DOI: 10.1145/3627673.3679630.
You can read the full text:
Resources
Contributors
The following have contributed to this page
